Keep It Simple: Multi-Key Episodic Memory Retrieval for Ultra-Long Video Understanding
Abstract
MERIT improves ultra-long video understanding by prioritizing high-recall episodic retrieval and deferring temporal relation reasoning to inference-time expansion.
When videos extend from hours to days, directly processing them end-to-end becomes impractical for current Multi-modal Large Language Models (MLLMs). This ultra-long setting necessitates a two-stage paradigm: query-agnostic memory construction followed by retrieval-based inference. Prior work invests in complex memory construction to pre-model high-level relations in videos, despite not knowing the downstream query at build time. We instead prioritize high-recall retrievability during memory building, and defer query-specific, high-level relation composition to inference time. To this end, we propose MERIT(Multi-key Episodic Retrieval with Inference-time Temporal expansion), a simple yet effective agentic framework for ultra-long video understanding. First, we formulate an episodic multi-key representation that enables precise retrieval of fine-grained memories through a simple key-matching mechanism. Second, we introduce a neighbor filtering mechanism to capture broader semantic context without the massive computational overhead of global memory construction. This is achieved by expanding the temporal scope exclusively around the retrieved segments at inference time. By leveraging simple key-matching with this on-demand temporal expansion, MERIT achieves state-of-the-art performance across three long-video benchmarks: EgoLifeQA, LVBench, and Video-MME (Long).
Community
We introduce MERIT(Multi-key Episodic Retrieval with Inference-time Temporal expansion) for ultra-long video QA. Prior work builds heavy memory before knowing the query — recursive summaries, multi-scale knowledge graphs, continuous consolidation. We argue that's the wrong place to spend intelligence. MERIT keeps memory simple and defers semantic composition to query time.
Two components:
- Multi-key indexing. Each 30s clip is dense-captioned once, then indexed by four complementary keys (event, dialogue, object, summary). A clip is scored by its best-matching key, so queries with different intents can all find it.
- Neighbor filtering. At inference, we expand ±1 min around each retrieved clip and let the solver filter for query-relevant information — temporal context on demand, not precomputed.
Results: EgoLifeQA 71.2% (+5.6), LVBench 71.8% (+9.9), Video-MME(Long) 77.7%. Memory construction uses 3.8× fewer LLM calls and 8.4× fewer input tokens than the prior graph-based SOTA.
Get this paper in your agent:
hf papers read 2608.07663 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper